AgenticRoboRL, described in a new arXiv preprint, adapts the post-training techniques used for coding agents to robot planners. Rather than swapping in a more powerful closed-source planner or adding external memory, the authors train the planner itself. The recipe combines supervised fine-tuning with multi-turn reinforcement learning, and harness-specific adapters let one RL pipeline work across different robot execution mechanisms, including VLA skill invocation and control-code generation.
On LIBERO-PRO, the framework raised a 9B-parameter planner's success rate from 18.1% to 71.0% without external memory. That approaches the 72.1% achieved by Codex (GPT-5.5) with external memory. The authors say the results demonstrate post-training as an effective approach for improving robot planners, and they plan to release the framework and trained checkpoints.
This digest is based solely on the preprint; no independent or conflicting sources were provided.